🎨 TailorRec: From User Behavior to Creative Content

---

🚨 Current Personalized Content Generation Challenges:

• ❌ Generic Outputs - Most AI content generators produce one-size-fits-all content without understanding individual preferences
  
• ❌ Single Modality - Tools typically generate only text OR images OR videos
• ❌ Platform Isolation - No integration with real social platforms for authentic user understanding
  

💡 Personalized-AIGC Solution:

🎯 Recommendation Engine, Multimodal Analyzer, Profile Generator, and Content Creator All-in-One! We're building a future where AI truly understands YOU. 💡 Simply connect your social accounts. Personalized-AIGC autonomously analyzes your behavior, builds your preference profile, and generates tailor-made content—videos, posts, and more! 🚀

---

📑 Table of Contents

• 💡 Key Features
• 🏗️ Architecture
• 🚀 Quick Start
• 📦 Modules
• 📊 Evaluation

---

💡 Key Features

📊 Behavior2Profile Automatic Preference Learning Transform raw user behavior data from social platforms into comprehensive preference profiles through multimodal LLM analysis of videos, images, and text.

🎬 Profile2Video AI Video Production Studio Generate personalized video content including crosstalk, talk shows, music videos, meme edits, and news summaries tailored to user interests.

📱 Profile2Post Publication-Ready Posts Create engaging social media posts with AI-generated copy, images, and platform-specific links for Bilibili, Xiaohongshu, Zhihu, Douyin, and Weibo.

🔮 SSL-RecSys 30+ SOTA Models Leverage self-supervised learning recommendation with LightGCN, SGL, BERT4Rec, and more for precise user preference prediction.

---

🎯 End-to-End Personalized Content Pipeline

The Challenges We Solve:

Challenge	Traditional Approach	Our Solution
🎭 User Understanding	Manual input of preferences	Automatic behavior analysis from social platforms
🧠 Content Ideation	Generic prompts	Profile-driven creative idea generation
🎬 Video Production	Single clips, no consistency	Multi-agent studio with 7 video types
📱 Social Posts	Text-only generation	Rich multimedia with platform links
🔄 Real-time Adaptation	Static profiles	Dynamic profiling from live data

---

🔥 Why Personalized-AIGC?

🧠 Zero Manual Input	🎬 Multi-Modal Output	🌐 Platform Native	🤖 SOTA RecSys	⚡ Real-Time
Automatic Profiling	Video + Text + Image	Bilibili, Xiaohongshu, Hupu	30+ SSL Models	Live Data Analysis
No need to describe yourself—we learn from your actual behavior across platforms	Generate complete video productions, social media posts with images, and cross-platform content	Native integration with major Chinese social platforms for authentic user data	Powered by SSLRec (WSDM'24) with LightGCN, BERT4Rec, SGL, and more	Continuously update profiles with fresh user interaction data

---

🏗️ Architecture

📊 System Overview

Personalized-AIGC is a multi-agent content generation framework that transforms user behavioral data into personalized creative content. Our system seamlessly bridges recommendation systems with AIGC, allowing AI to truly understand and create for individual users.

🕷️ DATA COLLECTION LAYER 📺 Bilibili Spider • 📕 Xiaohongshu Spider • 🏀 Hupu Spider • 🎬 Douban Spider
🔮 RECOMMENDATION ENGINE (SSLRec) LightGCN • SGL • SimGCL • BERT4Rec • DuoRec • 30+ Models
🎥 VIDEO ANALYST Frame Extraction • Scene Understanding • Content Summarization		🖼️ IMAGE ANALYST Visual Recognition • Style Analysis • Object Detection
🧠 PROFILE GENERATION Item Profiles • User Profiles • Preference Ranking • Creative Ideas
🎬 VIDEO AGENT Cross Talk • Talk Show • MAD/SVC • Rhythm Edit • News • Commentary		📱 TEXT-IMAGE GENERATOR AI Copywriting • Image Generation • Platform Links • HTML Posts
✨ OUTPUT LAYER 🎬 Personalized Videos • 📱 Social Media Posts • 🖼️ Generated Images • 📊 User Insights

---

🎬 VideoAgent: Multi-Agent Video Production

🎭 Cross Talk Convert audio into Chinese crosstalk duo performances

🎤 Talk Show Transform content into stand-up format

🎵 MAD/SVC Voice conversion with custom lyrics

🔊 MAD/TTS Character dialogue adaptation

🎬 Rhythm Automatic beat-synced cuts

📰 News Event summarization videos

📖 Commentary Book to video narration

Built with cutting-edge tools:

🗣️ CosyVoice     →  Voice Cloning & TTS
🐟 Fish-Speech   →  Multilingual Speech Synthesis
🔗 ImageBind     →  Multimodal Embedding
🎧 Demucs        →  Audio Source Separation
🎼 Whisper       →  Speech Recognition

---

🚀 Quick Start

🖥️ Environment

OS: Linux, Windows
Python: 3.10+
CUDA: 12.1+ (recommended for GPU acceleration)

📥 Clone and Install

git clone https://github.com/your-username/Personalized-AIGC.git
cd Personalized-AIGC

# Create virtual environment
python -m venv paigc_env
source paigc_env/bin/activate  # Linux/Mac
# or
.\paigc_env\Scripts\activate   # Windows

# Install dependencies
pip install -e .

⚙️ Configuration

Create a .env file in the project root:

# Dataset: bilibili, redbook, hupu, douban
DATASET=bilibili

# LLM Configuration
CHAT_API_KEY=your_openai_api_key
CHAT_BASE_URL=https://api.openai.com/v1
CHAT_MODEL=gpt-4o

# Image Generation (Optional)
IMAGE_API_KEY=your_image_api_key
IMAGE_BASE_URL=https://generativelanguage.googleapis.com/v1beta
IMAGE_MODEL=gemini-2.0-flash

🎯 Usage

Run the complete pipeline with a single command:

python main.py

This will:

1. 🔮 Run SSL recommendation model
2. 🧠 Analyze user content (video, image, text)
3. 📊 Generate user profiles
4. ✨ Create personalized content

---

📦 Modules

1️⃣ Data Collection

from tools.bilibili_realtime import BilibiliRealTime

# Collect user favorites from Bilibili
bilibili = BilibiliRealTime()
bilibili.process_all_users(save_data=True, download_videos=True)

2️⃣ Recommendation Engine

from tools.sslrec import run_sslrec

# Run self-supervised recommendation
run_sslrec()

3️⃣ Multimodal Analysis

from tools.analyze import analyze

# Analyze all users' content with concurrent processing
analyze(max_workers=5)

4️⃣ Content Generation

from tools.product import product

# Generate personalized video and text-image content
product()

---